How AI makes video now
We went through all 1,546 videos in our log to see how many were made with AI, and how. Of the 453 we found, most did not come from a video generator. They were built with code.
Why code beats footage
Say “AI video” and most people picture a model conjuring footage from a text prompt. In our log, only 73 of the 453 AI-made videos use generated footage. 345 use visuals drawn by code that a language model wrote: p5.js sketches pulsing to a beat, Three.js scenes, Blender scripts, Remotion timelines, or plain HTML animated frame by frame. (One video can use several techniques, so the counts overlap; the rest use AI for music, voice or the script.)
Code simply offers more control. A language model can write an exact, repeatable animation, tweak it on request and sync it to music down to the frame. Video generators produce pretty clips that are hard to change. Among the creators we track, getting the most out of AI video right now means programming it.
One prompt, then out of the way
Human involvement is often minimal. Of the 377 videos where AI did most or all of the work, about two thirds describe the human’s part mainly as prompting: write a request, iterate a few times, publish. Others edited the output, picked the best takes or wrote their own lyrics, but many of the most watched pieces started from a single prompt.
At the far end are autonomous agent runs. In 55 videos an AI agent ran the production itself for hours, writing the script, calling other tools for music and images, assembling the cut and rendering the file. Several of the best known were twelve-hour runs by Claude Opus 5.5 working in Claude Code, with the human mostly watching.
The stack behind a lot of it
The most popular recipe pairs code-drawn visuals with music, a script and a voice. Typically a language model (most often Claude Opus 5.5) writes the lyrics and the visual code, a music generator such as Suno makes the song, and a voice tool such as ElevenLabs reads or sings. This is the stack behind the “Claude Pop” music videos that filled timelines in September and October 2026.
When generated footage does appear, the tools named most are Seedance 2.5 and Higgsfield, often directed by a language model that writes the prompts and plans the shots.
Read the numbers with care
Our log is not a random sample of the internet. We track videos about AI news and deliberately collect AI-made work, so both AI video and Claude are overrepresented, and the September surge in the month-by-month table partly reflects when we started collecting. We also did not re-watch every video: a model classified them from our descriptions and the creators’ own notes, and each video’s page shows what it found and how confident it was. A counted technique means it appears somewhere in the video, not that it made the whole thing. Another 261 videos only show something an AI built or did, such as an app, a game or a robot at work; they are counted separately from the 453 made with AI.
Still, the log captures how the craft works in late 2026: which techniques creators actually reach for, the tools they name, and how little of the work many of them now do by hand.
The techniques, one by one
Counts are videos in our log where the technique was used to make the video (one video can use several). Examples are the most viewed; some are small uploads that are not in search engines but are kept here as evidence.
Subject built by AI 382 videos
Most named: Claude Opus 5.5 (214), GPT-6 Astra (60), Claude Code (59), Gemini 4 Argon (46), Claude Fable 5.1 (37), Claude Sonnet 5.5 (37)
- code-generated interactive builds and simulations (Meet Kimi K3, 26,002,487 views)
- Gemini 4 Argon coded the interactive Monochord audiovisual web app in Canvas from a single user prompt (Turn Your Webcam Into a Playable Synthesizer with Gemini 4 Argon, 17,184,212 views)
- AI voice cloning of Uncle Roger demonstrated live and in dialogue (Fish Audio Drama 3 is Here, 15,269,925 views)
- GPT-6 Astra generated the 3D Unreal Engine simulation and cooperative survival agents (GPT-6 Astra builds an Unreal Engine world of Astra-powered people who start talking to each other, 11,786,215 views)
Visuals from AI-written code 345 videos
Most named: Claude Opus 5.5 (268), Claude Code (34), Suno (30), GPT-6 Astra (22), Claude Fable 5.5 (22), ElevenLabs (18)
- generating Three.js and TypeScript 3D playable worlds from sketches and prompts (Meet Kimi K3, 26,002,487 views)
- single prompt generated code rendering each frame (Western civilization in 2 minutes 16 seconds, 15,200,735 views)
- Gemini 4 Argon generated Blender Python scripts and 3D pagoda complex scene (Gemini 4 Argon: 3D World in Blender, 10,669,403 views)
- Gemini 4 Argon wrote procedural code for real-time visual animations and styles (Building Real-Time Animation with Gemini 4 Argon, 10,197,991 views)
Music 151 videos
Most named: Claude Opus 5.5 (117), Suno (38), ElevenLabs (17), Claude Code (15), Suno v6 (13), Claude (unspecified) (13)
- Lyria 3.5 generated music tracks used as audio demonstrations throughout the video (Introducing Lyria 3.5 Music Generation in Gemini, 41,239,627 views)
- AI techno-pop song generation and vocal synthesis (“Gave Opus 5.5 donald’s prompt, Midjourney, and a moodboard. 12 hours later, woke up to this”, 20,211,750 views)
- AI-composed musical score (Western civilization in 2 minutes 16 seconds, 15,200,735 views)
- AI pop song and vocal generation (“The Clodyssey”, 12,515,691 views)
Script or lyrics 119 videos
Most named: Claude Opus 5.5 (89), Suno (17), Claude Code (14), Seedance 2.5 (13), ElevenLabs (12), Claude (unspecified) (9)
- Autonomous agent generation run by Claude Opus 5.5 over 12 hours (“Gave Opus 5.5 donald’s prompt, Midjourney, and a moodboard. 12 hours later, woke up to this”, 20,211,750 views)
- single prompt generated code rendering each frame (Western civilization in 2 minutes 16 seconds, 15,200,735 views)
- Autonomous 12-hour agent run by Claude Opus 5.5 (“The Clodyssey”, 12,515,691 views)
- LLM-generated TypeScript and three.js/WebGL deterministic renderer (i’m upping my p(doom), 2,893,749 views)
Generated footage 73 videos
Most named: Claude Opus 5.5 (32), Seedance 2.5 (28), Higgsfield (11), Suno (8), GPT-6 Astra (6), Seedance 2.0 (5)
- Multimodal video generation and video-to-video style transfers (Introducing Gemini Omni: Create Anything from Anything, 26,045,325 views)
- Generative video generation for cinematic scenes and lip-sync (“The Clodyssey”, 12,515,691 views)
- Base video shots generated with Seedance 2.5 (I’m upping my P(doom) - Opus 5.5 (et al.), 4,059,347 views)
- early text-to-video generation clips of Will Smith eating spaghetti (Will Smith Eating Spaghetti AI Video - (2023 vs 2024), 2,663,581 views)
Images 46 videos
Most named: Claude Opus 5.5 (29), Seedance 2.5 (13), Higgsfield (7), Midjourney (6), GPT Image 2.5 (5), Claude Code (5)
- Midjourney image generation based on a moodboard (“Gave Opus 5.5 donald’s prompt, Midjourney, and a moodboard. 12 hours later, woke up to this”, 20,211,750 views)
- Midjourney v4 text-to-image portrait generation (Harry Potter by Balenciaga, 14,813,479 views)
- Nano Banana generated 2D reference concept imagery (Gemini 4 Argon: 3D World in Blender, 10,669,403 views)
- Character and style sheets generated with image models via fal (I’m upping my P(doom) - Opus 5.5 (et al.), 4,059,347 views)
Voice 139 videos
Most named: Claude Opus 5.5 (85), Suno (24), ElevenLabs (21), Claude Code (16), Suno v6 (9), Claude (unspecified) (9)
- AI techno-pop song generation and vocal synthesis (“Gave Opus 5.5 donald’s prompt, Midjourney, and a moodboard. 12 hours later, woke up to this”, 20,211,750 views)
- AI voice cloning of Uncle Roger demonstrated live and in dialogue (Fish Audio Drama 3 is Here, 15,269,925 views)
- ElevenLabs synthetic voice cloning (Harry Potter by Balenciaga, 14,813,479 views)
- AI pop song and vocal generation (“The Clodyssey”, 12,515,691 views)
Avatar or lip-sync 24 videos
Most named: Claude Opus 5.5 (11), GPT-6 Astra (5), ElevenLabs (4), Claude Fable 5.1 (2), Claude Code (2), Suno (2)
- D-ID talking-head facial animation and lip-sync (Harry Potter by Balenciaga, 14,813,479 views)
- Animated anime idol VTuber avatar lip-synced to the track (Nothing Went Foom!, 937,024 views)
- synthetic presenter generation with lip-sync (“High-end Netflix-style documentary about superintelligence for normies”, directed by an Opus 5.5 agent via Runway MCP, 913,535 views)
- Synthetic voice generation with synchronized lip motion (What Remains | Short Film | Finalist · Seoul International AI Film Festival 2026, 803,150 views)
Agent-run production 55 videos
Most named: Claude Opus 5.5 (38), Claude Code (13), Suno (9), Higgsfield (9), Seedance 2.5 (9), GPT-6 Astra (6)
- Autonomous agent generation run by Claude Opus 5.5 over 12 hours (“Gave Opus 5.5 donald’s prompt, Midjourney, and a moodboard. 12 hours later, woke up to this”, 20,211,750 views)
- Autonomous 12-hour agent run by Claude Opus 5.5 (“The Clodyssey”, 12,515,691 views)
- Opus 5.5 running autonomously in Claude Code for 12 hours from a single prompt (I’m upping my P(doom) - Opus 5.5 (et al.), 4,059,347 views)
- Claude Opus 5.5 and subagents wrote storyboard, animation guide, and p5.js/p5.brush scene code (Claude Pop - I’m Upping My P(Doom), 3,095,861 views)
Editing, captions, translation 24 videos
Most named: Claude Opus 5.5 (14), Suno (7), Claude Code (6), Higgsfield (4), GPT-6 Astra (4), Seedance 2.5 (4)
- Claude edited CapCut draft files directly to sync clips and keyframes (Claude Edited My Entire CapCut Video (I Didn’t Touch the Timeline), 699,776 views)
- Claude Code orchestrated automated video cutting, audio insertion, and motion graphics (I Fully Automated My Video Editing Using Claude Code (Full Walkthrough), 672,095 views)
- GPT-6 Astra orchestrated end-to-end video production autonomously from a single prompt (GPT-6 Astra Made This Entire Video, 459,245 views)
- Suno generated the original music track (Opus 5.5 drives After Effects: skeleton-tracked kinetic type on an AI dance video, 381,456 views)
Research and checking 9 videos
Most named: Claude Opus 5.5 (3), ChatGPT (2), Claude Code (2), ElevenLabs (2), HeyGen (2), Gemini 4 Argon (1)
- ChatGPT-built ‘Podcast Studio’ web app used to gather research and generate interview questions (OpenAI Co-Founder: Start Building With AI Before You Feel Ready | Greg Brockman, 472,555 views)
- GPT-6 Astra orchestrated end-to-end video production autonomously from a single prompt (GPT-6 Astra Made This Entire Video, 459,245 views)
- Claude wrote the script and conducted computational linguistics analysis (“Pick a problem or mystery that obsesses you”: Claude tries the Voynich Manuscript and makes a movie about it, 243,861 views)
- Claude Fable 5 ran an autonomous video production agent via Claude Code’s /goal command (Claude Fable 5 Made This Entire Video By Itself., 155,686 views)
AI playing or controlling 30 videos
Most named: GPT-6 Astra (6), Claude Opus 5.5 (5), Claude Sonnet 5.5 (4), Gemini 4 Argon (3), Gemini Robotics 2 (3), Claude Fable 5.1 (2)
- autonomous block-by-block construction via API calls in Minecraft (“Build what you think you look like”: Claude Sonnet 5.5 vs GPT-6 Astra, 7,257,834 views)
- Gemini Robotics 2 foundation model autonomously controlling humanoid and bimanual robots in real time (Gemini Robotics 2 brings whole body intelligence to robots, 5,941,425 views)
- autonomous humanoid robot control by foundation model policy (Unitree General-Purpose Humanoid Foundation Model Fully Upgrade Major Open Source, 4,532,698 views)
- GPT-6 Astra and Claude Fable 5.1 played a full chess game against each other (GPT-6 ASTRA vs CLAUDE FABLE 5.1 (INSANE CHESS MATCH), 2,998,785 views)
The usual recipes
The most common combinations of techniques in one video.
- Subject built by AI + Visuals from AI-written code 97
- Visuals from AI-written code 72
- Visuals from AI-written code + Music + Script or lyrics + Voice 25
- Visuals from AI-written code + Music 22
- Visuals from AI-written code + Music + Script or lyrics 21
- Visuals from AI-written code + Music + Voice 19
- Voice 11
- Generated footage 11
The tools people name
| Tool or model | Videos |
|---|---|
| Claude Opus 5.5 | 303 |
| Claude Code | 41 |
| Suno | 38 |
| GPT-6 Astra | 29 |
| Seedance 2.5 | 29 |
| ElevenLabs | 24 |
| Claude (unspecified) | 23 |
| Claude Fable 5.5 | 22 |
| Claude Sonnet 5.5 | 17 |
| Gemini 4 Argon | 16 |
| Higgsfield | 16 |
| Claude Fable 5.1 | 16 |
| Suno v6 | 13 |
| GPT Image 2.5 | 6 |
| Claude Haiku 5.5 | 6 |
| Midjourney | 6 |
| Claude Fable 5 | 6 |
| Seedance 2.0 | 5 |
| Udio | 5 |
| Higgsfield MCP | 5 |
| GPT-6 Sol | 4 |
| Sora | 4 |
| ChatGPT | 3 |
| Fish Audio Drama 3 | 3 |
| Nano Banana | 3 |
Month by month
| Month | Videos logged | Mostly or fully AI-made |
|---|---|---|
| January 2026 | 5 | 3 (60%) |
| February 2026 | 6 | 2 (33%) |
| March 2026 | 11 | 2 (18%) |
| April 2026 | 9 | 3 (33%) |
| May 2026 | 39 | 4 (10%) |
| June 2026 | 18 | 4 (22%) |
| July 2026 | 46 | 8 (17%) |
| August 2026 | 37 | 4 (11%) |
| September 2026 | 738 | 324 (44%) |
| October 2026 | 620 | 110 (18%) |
Generative models in our registry
Video, image, music and world models, newest first within each kind, each linked to how to use it. The full registry also covers language, speech and robotics models: all models.
Video 18
- FLUX 3 Max Black Forest Labs
Fast frontier-quality video; Post-trained FLUX 3 Video by fal Research; Keyframes-to-video - Kling 4.0 (and 4.0 Flash) Kuaishou
30-second single-pass clips; Omni Reference with keyframes - Tavus Griffin (Human Interaction Model) Tavus
Full-duplex face-to-face video conversation; 'Video Turing test' result (vendor study) - HeyGen Video 1 HeyGen
Whole scene with sound in one pass; Text, image and reference modes; Up to 30 s and 2K (added after launch) - Utopai X Utopai Studios
Production-aware agentic video model; Reflections, caustics and spatial consistency - Wan 3.0 Alibaba Cloud (Tongyi Lab)
30-second single-pass clips with native audio; Documents and web pages as input; Multi-reference generation and editing - LTX-2.5 Lightricks, open weights
Open audio-video generation with multishot consistency; Fast generation; Robotics / physical-AI checkpoint - Dreamina Seedance 2.5 ByteDance Seed
30-second single-shot generation with extensions; 50 multimodal references in one pass; Timestamp-level audio/video editing - MiniMax H3 MiniMax, open weights
Open omni-modal video model with native audio; H3-Context-IR prompt pipeline; 768P to 2K regeneration - FLUX 3 Black Forest Labs
Unified image/video/audio/action model; Video with native synced audio; FLUX 3 Action for robotics - Luma Ray3.2 Luma AI
Multi-keyframe direction; Native HDR with 16-bit EXR export; Multi-face performance tracking and reframe - Grok Imagine Video 1.5 xAI
Image-to-video up to 15 s; Per-second pricing, text or image input - Runway Aleph 2.0 Runway
In-context video editing of real footage; Edit one frame, propagate to the clip - Gemini Omni Flash (Omni 1.1 Flash) Google DeepMind
Any-input video generation; Conversational video editing; Up to 4K output - Dreamina Seedance 2.0 ByteDance Seed
Unified audio-video joint generation; Multimodal reference and editing; Technical report - Kling 3.0 (VIDEO 3.0 / 3.0 Omni) Kuaishou
Multi-shot storyboards; Native multilingual audio; Unified Omni model with element consistency - Runway Gen-4.5 Runway
#1 on Artificial Analysis text-to-video at launch; HDR and professional output formats - Veo 3.1 Google DeepMind
Native audio in every clip; Reference images and first/last frame control; Video extension to a minute+
Image 26
- FLUX 3 Image Klein Black Forest Labs, open weights
Open-weights 4K generation with bounding boxes; Consumer-GPU builds - Krea 3 (Fast / High / Ultra modes) Krea
Ultra thinking mode; Pixel-exact regional edits; Bounding boxes with per-region notes - Nano Banana 2.1 (Gemini Nano Banana 2.1) Google DeepMind
Half the per-image price of Nano Banana 2; Built on Gemini 3.6 Flash; Configurable thinking and search grounding - Qwen-Image-2.1 Alibaba (Qwen), open weights
Native RGBA generation and editing; Multi-reference editing - GPT Image 2.5 Flare OpenAI
Fast everyday image generation; Inpainting - GPT Image 2.5 Sunburst OpenAI
Most capable OpenAI image model; Replacement for gpt-image-1.5/1-mini - MAI-Image-2.6-Flash Microsoft
Low-cost, fast tier of MAI-Image-2.6; Bing web grounding and auto aspect ratio; Up to five reference images for edits - MAI-Image-2.6 Microsoft
Bing web grounding; Model-chosen aspect ratio; Multi-image reference editing - Grok Imagine Image 2.0 xAI
Generation + editing in one model; Top-2 on Arena image leaderboards at launch - Midjourney V8.2 Midjourney
Instruction-based edit model; Improved personalization; Rewritten V8 core with native 2K and better text - Qwen-Image-3.0 (Pro) Alibaba (Qwen)
Dense single-pass layouts; Tiny, multilingual text rendering; Closed release (break from open Qwen-Image) - Seedream 5.0 Pro ByteDance Seed
Region-precise editing with drawn boxes; Layer separation; Native text in 14 languages and dense layouts - Muse Image Meta
Agentic image generation; Test-time compute scaling; Multi-reference composition and multi-turn editing - Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image) Google DeepMind
Lowest-cost Gemini image model; Video-as-input image generation - Ideogram 4.0 Ideogram, open weights
Structured JSON prompting with layout control; Best-in-class multilingual text rendering; First Ideogram open-weight model - Recraft V4.1 Recraft
Native vector (SVG) generation; Utility variant for mockups; V4.1 Flash - GPT Image 2 OpenAI
Batch image generation; DALL-E replacement - Nano Banana 2 (Gemini 3.1 Flash Image) Google DeepMind
Pro quality at Flash speed; Image search grounding; Text rendering and in-image translation - FLUX.2 [klein] (4B / 9B) Black Forest Labs, open weights
Sub-second generation and editing; Apache-2.0 open weights (4B); KV-cached 9B variant - FLUX.2 [max] Black Forest Labs
Grounded generation with web search; Highest editing consistency in FLUX.2 - FLUX.2 [dev] Black Forest Labs, open weights
32B open-weight generation + multi-reference editing; VLM-conditioned rectified flow transformer - FLUX.2 [pro] Black Forest Labs
Multi-reference editing (up to 10 images); 4MP editing and production-grade typography - Nano Banana Pro (Gemini 3 Pro Image) Google DeepMind
Accurate multilingual text in images; Search-grounded visuals; Multi-image composition - Nano Banana (Gemini 2.5 Flash Image) Google DeepMind
Conversational image editing; Multi-image fusion and targeted edits - FLUX.1 Kontext [pro] / [max] Black Forest Labs
In-context image editing; Open-weight editing sibling - Stable Diffusion 3.5 Large Stability AI, open weights
Open MMDiT weights, free for small businesses; Broad hardware optimization
Music 12
- Eleven Music v2.5 ElevenLabs
Commercially cleared music generation; Composition plans and audio reference; Composition-plan chunks via API - YuE2-3B Multimodal Art Projection (M-A-P), open weights
Score-first song generation; Zero-shot covers and agentic editing - Suno v6 (v6, v6-wild, v6-mini) Suno
Trained only on licensed music; Three-variant lineup; Natural-language section and lyric editing - Lyria 3.5 Google DeepMind
Full songs with vocals and lyrics; Image-conditioned music; SynthID-watermarked audio - MiniMax Music 3.0 MiniMax, open weights
Open-weights full songs up to ~5 minutes in one pass; Hierarchical Global/Local LLM with continuous hidden-state synthesis; Consumer-GPU inference - Mureka V9.5 (and O3) Kunlun Tech (Skywork AI)
MusiCoT (music chain-of-thought) planning; MuCo creation agent; Fine-tuning API and vocal cloning - Stable Audio 3.0 Stability AI, open weights
Tracks over 6 minutes; Fully licensed training data; On-device small models - SongGeneration 2 (LeVo 2) Tencent AI Lab, open weights
Hybrid LLM-diffusion full songs up to 4:30; Hierarchical semantic planning + track-specific refinement - Lyria 3 (Clip / Pro) Google DeepMind
Songs with vocals and auto-written lyrics in the Gemini app; SynthID watermark + detection in Gemini; Full songs with structure control (Lyria 3 Pro) - ACE-Step 1.5 (incl. 1.5 XL) ACE Studio & StepFun, open weights
Full songs in seconds on consumer hardware; LM planner + DiT synthesizer; Editing and personalization toolkit - Lyria 2 Google DeepMind
Instrumental clips with negative prompting - Lyria RealTime Google DeepMind
Interactive streaming music generation; Live musical parameters
World models 5
- Xiaomi-Robotics-U0 (38B) / U0-4B Xiaomi, open weights
Unified embodied synthesis; Data engine for VLAs; FlashAR fast decoding - Cosmos 3 (Nano / Super) NVIDIA, open weights
Unified omni world model (generation + reasoning + action); Open omnimodal I/O; Leaderboard results - 1X World Model (1XWM) 1X Technologies
Video world model used as the robot policy; Learns from human egocentric video; World model for policy evaluation - Cosmos Predict 2.5 / Transfer 2.5 NVIDIA, open weights
Unified Text2World / Image2World / Video2World; Multi-control world-to-world transfer - Genie 3 Google DeepMind
Real-time interactive world generation; World memory / consistency; Promptable world events
How this was made: every video in the log has a description written by Gemini from the video itself, plus our notes, often taken from the creators' own posts. Gemini 3.8 Flash then read those texts (not the videos again) and classified how AI was used to make each one, in eleven fixed categories; it was told that showing or discussing an AI product does not count. The classification is shown on each video's page. Counts update with every build; last 25 October 2026.